A real-time feeding behavior monitoring system for individual yak based on facial recognition model

Feeding behavior is known to affect the welfare and fattening efficiency of yaks in feedlots. With the advancement of machine vision and sensor technologies, the monitoring of animal behavior is progressively shifting from manual observation towards automated and stress-free methodologies. In this study, a real-time detection model for individual yak feeding and picking behavior was developed using YOLO series model and StrongSORT tracking model. In this study, we used videos collected from 11 yaks raised in two pens to train the yak face classification with YOLO series models and tracked their individual behavior using the StrongSORT tracking model. The yak behavior patterns detected in trough range were defined as feeding and picking, and the overall detection performance of these two behavior patterns was described using indicators such as accuracy, precision, recall, and F1-score. The improved YOLOv8 and Strongsort model achieved the best performance, with detection accuracy, precision, recall, and F1-score of 98.76%, 98.77%, 98.68%, and 98.72%, respectively. Yaks which have similar facial features have a chance of being confused with one another. A few yaks were misidentified because their faces were obscured by another yak's head or staff. The results showed that individual yak feeding behaviors can be accurately detected in real-time using the YOLO series and StrongSORT models, and this approach has the potential to be used for longer-term yak feeding monitoring. In the future, a dataset of yaks in various cultivate environments, group sizes, and lighting conditions will be included. Furthermore, the relationship between feeding time and yak weight gain will be investigated in order to predict livestock weight.